Foundry

Benchmarks

Measured against the frontier.

Two systems, one standard. Helix is scored on advanced reasoning benchmarks every frontier lab reports. The Vortex Engineering Module is scored on real engineering work, against frontier models and human experts.

The model

Helix

Our enhanced inference engine and system one orchestration model.

Helix decides how each problem gets solved: how much reasoning to spend, which tools and systems to bring in, and how to check the answer before it is returned. It is served as one model behind one API, so the gains arrive without changing how you build.

Enhanced inference
More correct answers for every dollar of compute.
System orchestration
Plans, delegates and coordinates multi-step work end to end.
Checked answers
Work is verified before it reaches you, so it can be trusted.

ResultsARC-AGI-2SimpleQA VerifiedHumanity's Last ExamEngDesign

The engineering module

Vortex Engineering

The harness, data and evaluation that turn a model into an engineer.

Everything a model needs to do engineering work in production, packaged as one module. Run it with Helix or bring your own model: the module is what makes the work reliable.

Vortex Harness
Agents that operate the CAD, CAE and solvers already in your program, with the workflows engineers use.
Engineering data
Geometry, meshes, simulations and requirements, connected into one model of the program.
Evaluation
Every result held to expert references and engineering checks before it is kept.
Deployment
On-premises or in your cloud, with access controls and a full audit trail.

ResultsCFD meshingOpen-weight models

At a glance

01Reasoning

ARC-AGI-2

Score against cost per task. Up and to the left is better. The dashed line is the best published score at each price.

Top ten by score

#ModelProviderScoreCost / task

Source: official ARC Prize leaderboard, arcprize.org/leaderboard (semi-private set). Helix is scored on the ARC-AGI-2 public evaluation set (120 tasks, pass@2) and is not an ARC Prize–verified entry. Cost axis is logarithmic.

02Factuality

SimpleQA Verified

Short-form factual accuracy on 1,000 questions, no tools. Whiskers are the published 95% intervals.

    Source: official SimpleQA Verified leaderboard by Google DeepMind on Kaggle, kaggle.com/benchmarks/deepmind/simpleqa-verified (updated Oct 20, 2026; top 17 models shown, plus Helix; whiskers are the published 95% intervals).

    03Expert knowledge

    Humanity's Last Exam

    Expert-level closed-ended questions, text only, no tools.

      Source: official Scale Labs leaderboard, labs.scale.com/leaderboard/humanitys_last_exam_text_only.

      04Engineering

      EngDesign

      Open-ended engineering design tasks across 9 domains: control systems, analog and digital circuits, structural and mechanical design, robotics, computer architecture and signal processing.

        Source: EngDesign benchmark, agi4engineering.github.io/Eng-Design. * Helix model running in the Vortex Harness.

        05Vortex Harness

        CFD meshing

        The Vortex Harness, our engineering harness, against frontier models on automated CFD meshing, each measured against a human reference mesh. Pass rate, mean boundary-layer coverage across 17 boundary-layer cases, and mean overall mesh score.

        Official benchmark details on the VortexAI blog

        The human reference is the expert mesh for each case; frontier models run on the same cases. Mesh score is on a 0 to 1 scale. Benchmark details: VortexAI: Automated CFD Mesh Generation Powered by LLM Agents.

        06Open-weight models

        Nemotron 3 Ultra, bare and in the harness.

        The same open-weight model on the same meshing cases, run on its own and inside the Vortex Harness. The tick on each track is the human reference.

        Official benchmark details on the VortexAI blog

        Cost is per successful mesh in US dollars; time is the mean per case in hours. Mesh score is on a 0 to 1 scale, as in CFD meshing above.

        Foundry

        Enhanced reasoning. Frontier intelligence.

        Tell us what your model needs to do.

        Talk to us